<h1 id="minimax-h3-amd-hip-and-multi-gpu-tips">MiniMax H3 — AMD HIP &amp; Multi-GPU Tips</h1>Here are a few tips to get MiniMax H3 running smoothly on AMD GPUs with HIP/ROCm, including a working multi-GPU setup.For reference, my setup is 2× AMD Radeon RX 7900 XTX, with 24 GB VRAM each. Some of the settings below, especially the multi-GPU VRAM allocation, are therefore specific to this configuration and may need to be adjusted for other GPUs.<h2 id="1.-use-the-patientx-cfz-rocm-build">1. Use the patientx-cfz ROCm build</h2>Use the build from patientx-cfz:<a target="_blank" rel="ugc" href="https://github.com/patientx-cfz/comfyui-rocm">https://github.com/patientx-cfz/comfyui-rocm</a>It works very well for me and has been stable and error-free. Follow the installation instructions on GitHub.After installing, make sure to update everything to the latest version using the included batch files:<pre><code>comfyui-rocm-updater.bat
rocm-pytorch-package-updater.bat
</code></pre>Then edit:<pre><code>comfyui-rocm.bat
</code></pre>Find the line:<pre><code>set PARAMS=
</code></pre>and add:<pre><code>--use-sage-attention
</code></pre>at the end.<hr /><h2 id="2.-download-the-q8_0-minimax-h3-models">2. Download the Q8_0 MiniMax H3 models</h2>I recommend using the Q8_0 models.The INT8 versions also work and are roughly as fast as Q8, but in my testing they require significantly more VRAM with no noticeable quality benefit.I used the Unsloth versions:<a target="_blank" rel="ugc" href="https://huggingface.co/unsloth/MiniMax-H3-GGUF/tree/main">https://huggingface.co/unsloth/MiniMax-H3-GGUF/tree/main</a>Specifically:<pre><code>minimax_h3_fl2va_pruned-Q8_0.gguf
minimax_h3_ref2va_pruned-Q8_0.gguf
</code></pre>If you only have one 24 GB GPU, I suggest starting with Q5_0 quantization instead.<hr /><h2 id="3.-use-the-recommended-clip-in-nvfp4">3. Use the recommended CLIP in NVFP4</h2>Download the recommended CLIP:<pre><code>qwen3vl_32b_minimax_h3_nvfp4_awq
</code></pre>This was the main point of confusion for me.I'm familiar with NVFP4 from the LLM world, and because NVFP4 is associated with NVIDIA hardware, I initially assumed it wouldn't work on AMD.I tried several GGUF CLIP models and managed to get them working only for T2V. For the other workflows, I had no luck.The important part is:You do NOT need a GGUF CLIP here. NVFP4 works on AMD in this setup.The included comfy-kitchen module handles the quantized model transparently, so the recommended NVFP4 CLIP can be used directly.<hr /><h2 id="4.-install-the-required-custom-nodes">4. Install the required custom nodes</h2>Using ComfyUI Manager, install:<pre><code>comfyui-multigpu
ComfyUI-GGUF
</code></pre>After installing <code>comfyui-multigpu</code>, you'll get several new loader nodes.For the GGUF model, use:<pre><code>UnetLoaderGGUFDisTorch2MultiGpu
</code></pre>On my 2× RX 7900 XTX 24 GB setup, I use:<pre><code>virtual_vram_gb = 20
</code></pre>This value is specific to my hardware configuration and should not be treated as a universal recommendation. If you're using GPUs with a different amount of VRAM, you'll likely need to adjust it accordingly.In my setup, this lets the second 7900 XTX act as a very fast cache/offload device for the primary 7900 XTX doing the main computation.<hr /><h2 id="5.-optional:-turbo-loras">5. Optional: Turbo LoRAs</h2>You can add LoRAs on top if you want.The MiniMax H3 LoRA ecosystem is evolving quickly, but these were among the first Turbo LoRAs available:<a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main">https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main</a>They allow you to generate with around 4 or 8 steps instead of the default 20, depending on the LoRA and workflow.<hr />This setup is currently working very well for me on 2× RX 7900 XTX 24 GB, especially with the multi-GPU loader.I'll attach the workflow I used as well.

Files with SHA256: 8443c46abd4683841720ab91373da052f887b8a973bc5a890a8018c6baa655eb